Coding has long been treated as the foundation of qualitative analysis, the necessary element that transforms raw data into meaningful insight.
Curated by Business Science Daily — peer-reviewed sources, human-verified.
Learn more
About Our Curation Process
Business Science Daily curates academic research in business and management. Each featured study is selected from reputable, peer-reviewed journals, institutional repositories, or working papers (e.g., Elsevier, Sage, NBER, SSRN).
Articles are carefully summarized to ensure clarity and accuracy, with direct citations or links to original sources. Our process emphasizes transparency, academic integrity, and accessibility for a broader audience.
Learn more in our Editorial Standards & AI Policy.
But what if it isn’t necessary anymore?
Large language models can read, summarise, and synthesise text at scale. As of now, they can retrieve relevant passages, compare perspectives across cases, and even suggest thematic connections.
Do we know if AI can analyse qualitative data without applying a single code? If yes, why are researchers still spending weeks or months segmenting their data?
Susanne Friese, a qualitative methodologist, has talked about the issue. In her recent article in Qualitative Inquiry, she argues that the coding paradigm is not just inefficient but epistemologically obsolete.
She argues that the problem is no longer about coding automation. The author proposes a new method that replaces coding with a structured playground for researchers and large language models. The result, she argues, is not a faster version of the same process but a fundamentally different landscape.
Is Automated Coding The Solution?
There are a growing number of studies that claim to automate qualitative coding using large language models. Fiese argues that many of these papers are authored by computer scientists, data scientists, or human-computer interaction researchers. They treat qualitative data as text to be segmented, labelled, or clustered.
The authors typically apply fewer than 10 predefined codes to short texts such as survey responses or tweets. The process bears closer resemblance to quantitative content analysis than to genuine qualitative inquiry.
True qualitative coding is something different, Friese says. Qualitative researchers typically work with full transcripts — often 10 to 35 interviews. They develop between 80 and 250 distinct codes, frequently arranged into hierarchical categories. Therefore, the process is iterative and deeply interpretive. Analysts shift between inductive openness and deductive structure, exploring contradictions, emotional tone, and subtle meaning. Finally, the goal is not just thematic description but interpretive depth, often culminating in conceptual categories, metaphors, or grounded theories.
The gap between these two approaches is not a matter of scale. Friese believes it is a matter of what counts as knowledge. When researchers treat qualitative analysis as a classification task, they flatten it. They sacrifice reflexivity, nuance, and methodological rigor in exchange for speed. And in doing so, they risk producing findings that are not only shallow but misleading.
Her critique is not that AI cannot assist qualitative research. It is that the way it is being used often misunderstands the very nature of the enterprise.
The Hybrid Trap
To understand the significance of Friese’s proposal, it helps to see the landscape she is navigating. Established CAQDAS platforms — MAXQDA, ATLAS.ti, NVivo — have all introduced generative AI features in recent years. But the platforms’ approach has been to integrate AI into existing coding-based workflows rather than replace them.
These hybrid solutions, Friese suggests, may represent a transitional phase rather than a destination. They speed up coding without questioning whether it is needed at all.
Conversational Analysis with AI
The core of Friese’s proposal is Conversational Analysis with AI. She argues that CA, that is powered by AI, or CA^AI, replaces coding with iterative dialogue between researcher and language model. The method is explained in five steps. Each step positions the researcher as an active analyst who guides the AI toward insight.
Conversational Analysis with AI replaces traditional coding with structured, iterative dialogue between researcher and AI. The method proceeds topic by topic, not document by document.
Getting to know the data
Generate summaries and extract preliminary themes with AI assistance. This serves as orientation — identifying key areas for exploration.
orientation · familiarisationPreparing for analysis
Select a topic and develop a set of exploratory questions. These become the analytical scaffolding for that topic — transparent, replicable, and topic‑by‑topic.
scaffolding · questionsAsking questions
Dialogic exchange with AI. Responses are openings for deeper exploration: follow‑ups, clarifications, case comparisons, or challenges to first impressions.
Abductive reasoning shines here — probing anomalies, brainstorming explanations.
dialogue · abductionSynthesizing insights
Slow down to engage deeply with the conversation. Write a synthesis grounded in theory and experience — or co‑develop it with the AI. The LLM acts as an epistemic actor.
synthesis · co‑constructionElevating the analysis
Optional but valuable. Move beyond description: identify relationships between themes, integrate findings into existing frameworks, or contribute new theoretical insights.
theory · abstractionThe process is iterative, not fixed. When researchers discover that something important was left out, the original questions can be revised and applied in the next loop. This allows the analysis to evolve in response to emerging findings.
Generate summaries and extract preliminary themes with AI assistance. This serves as orientation — identifying key areas for exploration.
The researcher begins by familiarising themselves with the material, using the AI to surface initial patterns and potential avenues for deeper inquiry.
Key activities: Summarising documents, identifying recurring concepts, noting unusual or unexpected content.
Select a topic and develop a set of exploratory questions. These questions become the analytical scaffolding for that topic and serve as a means of making the analysis transparent and replicable.
Crucially, CAAI proceeds topic by topic, rather than document by document. This contrasts sharply with conventional coding, where researchers typically seek diversity early on, coding across all topics simultaneously.
The dialogic exchange itself. The researcher engages the AI using the questions developed earlier.
AI responses are not treated as definitive answers — they are openings for deeper exploration: follow‑up questions, requests for clarification, comparisons across cases, or challenges to initial impressions.
Interpretation begins to take shape not as something delivered by the AI but as a process the researcher actively guides and participates in.
Abductive reasoning becomes particularly valuable here. Researchers can ask exploratory questions to probe anomalies, brainstorm plausible explanations, and iterate between data and hypotheses. The AI becomes a creative thinking partner, not a coding engine.
The researcher slows down to engage deeply with the conversation.
This can be done independently, by reading through the exchange and writing a synthesis grounded in theory and experience.
Or it can be co‑developed with the AI, by uploading the conversation and shaping a draft collaboratively. The LLM acts as an epistemic actor — not because it understands, but because it contributes distributed patterns of association drawn from its intertextual training.
Optional but valuable. This step involves elevating the analysis through explicit theoretical reasoning.
The objective is not merely to describe patterns but to engage in conceptual abstraction: identifying relationships between themes, integrating findings into existing theoretical frameworks, or contributing to new theoretical insights.
The researcher remains firmly in control, using the AI to test emerging interpretations, explore counter‑arguments, and refine the theoretical contribution.
What It Means And Why It Matters
Traditional qualitative inquiry has been grounded in the interpretive paradigm. In other words, knowledge is constructed through immersive engagement with empirical material.
Meaning does not reside in the data but is generated through a relationship between analyst, context, and text. As a result, knowing is deeply shaped by the knower’s positionality, experience, and reflexive engagement.
When AI is added to this equation, something shifts. The AI does not replace the researcher but becomes part of a triadic interpretive space.
The LLM’s outputs act as provocations that the researcher responds to, questions, and situates. Understanding becomes a process of navigating between perspectives: that of the participant, the AI model, and the researcher’s own evolving framework.
This raises a provocative possibility. Engaging with LLMs allows researchers to access perspectives that may otherwise remain inaccessible.
Generative AI challenges the traditional assumption that coding is intrinsic to qualitative research. Friese argues that coding — the systematic segmentation and labelling of text — can be replaced by structured dialogic interaction between researchers and large language models.
The method, Conversational Analysis with AI (CAAI), reframes analysis as iterative questioning, synthesis, and reflexive interpretation rather than categorization and retrieval.
Friese draws a critical distinction between true qualitative coding and what she calls classification proxies — surface-level labels assigned by AI that lack interpretive depth.
Iterative, interpretive, contextually grounded.
- 80–250 codes per project
- Hierarchical categories and subcodes
- Emergent insight through immersion, memoing, reconfiguration
- Dialogic, theory-informed
- Goal: interpretive depth, conceptual development
Shallow, pre-segmented, context-blind.
- Usually fewer than 10 codes
- Predefined categories, often from short texts
- Chunking destroys interpretive continuity
- Non-reproducible, overly reliant on surface phrasing
- Goal: speed, not depth
Established CAQDAS platforms have integrated generative AI features, but their approach has been cautious — integrating AI into existing coding-based workflows rather than replacing them.
Helps researchers apply codes segment by segment, providing explanations for each assignment. Serves as a second coder during pilot phases and a brainstorming aid. Researcher remains in control.
Full-scale automated AI-powered coding — but processes each paragraph in isolation. Leads to severe code proliferation: 400+ codes for two documents, 1,200+ for ten. Time spent cleaning often exceeds time saved.
CAAI replaces coding with structured, iterative dialogue between researcher and AI. The method proceeds topic by topic, not document by document.
Open-ended prompts allow patterns to emerge organically from the data.
Questions shaped by existing theoretical frameworks or prior findings.
Explore anomalies, contradictions — develop plausible explanations for surprising observations.
CAAI is grounded in a hermeneutic epistemology where knowledge is constructed dialogically — through interaction between researcher, text, and interpretive horizons. When AI is added, it becomes part of a triadic interpretive space.
- AI lacks Seinsverbundenheit: It does not experience, feel, or reflect in the ways human researchers do. Yet its outputs can still offer viable insights because it is trained on vast, socially-situated corpora.
- Epistemic expansion: AI can introduce interpretive angles, nuances, or contrasts that extend the scope of human inquiry.
- Interpretive triangulation: Not between data sources, but between human insight and machine-generated associations.
Traditional markers of rigor — coding frames, inter-coder agreement — no longer apply. CAAI replaces them with new strategies anchored in dialogic transparency and reflexive practice.
Use RAG-based platforms (e.g., QInsights) that ground outputs in source material. Avoid general-purpose chatbots — they risk hallucinations and lack traceability.
Set temperature between 0.2 and 0.5 for consistency. General-purpose chatbots operate at ~0.7, optimized for creativity, not research consistency.
The move away from coding brings significant gains but also important trade-offs. Recognizing these tensions is essential for ensuring that this paradigm shift remains grounded in interpretive and ethical commitments.
- Reorientation of analytic labour: From mechanical segmentation toward conceptual synthesis.
- Flexibility: Iterative exploration without rigid coding frames.
- Epistemic clarity: Questions, decisions, and syntheses are explicitly articulated.
- Transparency: Full audit trail through documented prompts and responses.
- Depth: Researchers can track meaning across documents and subgroups.
- Procedural grounding: Coding provides a clear structure and a sense of analytic discipline.
- Familiar validation tools: Inter-coder agreement, frequency counts, co-occurrence matrices.
- Collaborative anchors: Shared coding schemes provide a common reference point.
- Quantification: Frequency counts and code co-occurrence play a less central role.
- For researchers: Requires new competencies — asking the right questions, maintaining analytic focus, resisting the temptation to treat AI outputs as ready-made insights.
- For collaborative work: Interpretive alignment replaces procedural alignment. Researchers use the same question set, then compare syntheses.
- For mixed-methods: Quantification is not entirely absent — identifying which documents a theme appears in is feasible — but precise frequency counts remain unreliable with current LLMs.
Threats to Rigor?
Friese believes perhaps the most challenging aspect is the implications for reliability and validity. Traditional qualitative research has relied on coding frames and inter-coder agreement as markers of rigor.
CA^AI responds by reimagining what counts as rigor. Transparency is achieved through full documentation of the dialogic process. This includes the entire sequence of questions, AI responses, and researcher reflections, which is captured in the chat protocol.
This audit trail is more detailed than traditional coding systems, where the final code assignment often obscures the thought process behind it.
Validity is enhanced through traceability to source data. Friese recommends using retrieval-augmented generation systems that ground AI outputs in specific passages from the researcher’s own documents. This allows direct validation of responses against the empirical material.
Cross-person reliability can be assessed by having two researchers independently use the same set of questions and then comparing their syntheses. Temporal reliability can be checked by repeating the same analytic sequence after an interval. These strategies shift the focus from mechanical agreement to reflexive interpretive rigor.
The language model’s temperature setting plays a critical role. Lower temperatures reduce randomness, resulting in more predictable output. Friese recommends settings between 0.2 and 0.5 for research purposes, rather than the higher settings typically used by general-purpose chatbots.
The tensions, however, might arise. What is gained in flexibility and depth may be lost in procedural clarity and familiar validation tools. But Friese argues that the trade-off is worth making.
Pros & Cons
The author argues that moving away from coding brings significant gains. Researchers can spend less time organizing data into categories and more time interpreting meaning. The process becomes more flexible. This allows for iterative exploration without rigid coding frames. Analysis can also begin earlier, proceed more fluidly, and remain closer to the data’s context.
She added that epistemic clarity increases. Instead of collapsing interpretation into category counts, researchers articulate their questions, decisions, and syntheses explicitly. This creates a new form of transparency — where the analytic path can be traced through documented prompts, AI responses, and human reflections.
But something is also lost. Coding provides grounding. A clear procedure and a sense of analytic discipline. It externalizes thought and offers a tangible structure for managing complexity. For new researchers especially, coding offers a foothold. Letting go of it requires developing new competencies: asking the right questions, maintaining analytic focus, and resisting the temptation to treat AI outputs as ready-made insights.
Collaborative analysis may also feel less anchored without the familiar scaffolding of coding schemes. The procedural alignment of inter-coder agreement is replaced by interpretive alignment — researchers collaborating by posing the same questions and comparing syntheses. This lacks the numerical reassurance of agreement coefficients but introduces a more reflexive form of validation based on conceptual coherence and critical dialogue.
Friese acknowledges these tensions rather than dismissing them. The challenge, she suggests, lies in developing new practices that preserve the best of qualitative inquiry while embracing the epistemic possibilities made possible by dialogic AI.
The Final Thought
This is not the first time qualitative research has faced a technological disruption. The introduction of CAQDAS in the 1990s provoked similar anxieties about the mechanization of interpretation (see the University of Surrey’s CAQDAS Networking Project).
The challenge now is similar but more profound. Generative AI is not just a tool that accelerates existing processes. It is a technology that forces a reconsideration of what the processes are for. Whether that reconsideration will lead to a post-coding era — or a more sophisticated integration of AI with traditional practices — remains to be seen.